2019-05-28 13:12:27 +03:00
|
|
|
$ setconfig extensions.treemanifest=!
|
2017-04-18 04:08:07 +03:00
|
|
|
#require p4
|
|
|
|
|
|
2017-05-12 03:13:31 +03:00
|
|
|
$ . $TESTDIR/p4setup.sh
|
p4fastimport : introducing fast Perforce to Mercurial convert extension
Summary:
`p4fastimport` is a fast convert extensions for Perforce to Mercurial. It
is designed to generate filelogs in parallel from Perforce. It tries to
minimize the use of Perforce commands and reads from the the Perforce
store on a Perforce server directly.
The core of p4fastimport is the idea to generate a Mercurial filelog
directly from the underlying Perforce data, as a Perforce file in most
cases matches a filelog directly (per-file branches is an exception). To
generate a filelog we are reading each file for an imported revision. A
file in Perforce is locally either stored in RCS, as a compressed GZIP
or as an flat file (binaries). If we do not find a version locally on
disk we fallback to downloading it from Perforce.
We are generating manifests after all filelogs are imported. A manifest
is constructed by adding and removing files from an initial state. We
are generating the correct offset from a manifest into the filelog by
keeping track of how often a file was touched.
We then generate the changelog.
Linkrev generation is a bit tricky. For every file in Perforce know
to which changelist it belongs, as it's stored revisions contains the
changelist. E.g. 1.1422 is the file changed in the changelist 1422 (this
refers to the "original" changelist, before a potential renumbering,
which is why we use the -O switch). We use the CL number obtained
from the revision to reverse lookup the offset in the sorted list of
changelists, which corresponds to it's place in the changelog later,
and therefore it's correct linkrev.
Parallel imports: In order to run parallel imports we MUST keep one lock
at a time, even if we import multiple file logs at the same time. However
filelogs use a singular `fncache`, which will be corrupted if we generate
filelogs in parallel. To avoid this, repositories must be generated with
*fncache* disabled! This restricts `p4fastimport` with workers to run
only on case sensitive file systems.
Test Plan:
The included tests as well as multiple imports from a small testing
Perforce client. Afterwards successfully run `hg verify`
make tests
Reviewers: #idi, quark, durham
Reviewed By: durham
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4776651
Signature: t1:4776651:1492015012:0161c4f45eab4d3b64597d012188c5f2007e8f7d
2017-04-13 21:11:09 +03:00
|
|
|
|
|
|
|
|
populate the depot
|
|
|
|
|
$ mkdir Main
|
|
|
|
|
$ mkdir Main/b
|
|
|
|
|
$ echo a > Main/a
|
|
|
|
|
$ echo c > Main/b/c
|
|
|
|
|
$ echo d > Main/d
|
|
|
|
|
$ p4 add Main/a Main/b/c Main/d
|
|
|
|
|
//depot/Main/a#1 - opened for add
|
|
|
|
|
//depot/Main/b/c#1 - opened for add
|
|
|
|
|
//depot/Main/d#1 - opened for add
|
|
|
|
|
$ p4 submit -d initial
|
|
|
|
|
Submitting change 1.
|
|
|
|
|
Locking 3 files ...
|
|
|
|
|
add //depot/Main/a#1
|
|
|
|
|
add //depot/Main/b/c#1
|
|
|
|
|
add //depot/Main/d#1
|
|
|
|
|
Change 1 submitted.
|
|
|
|
|
|
|
|
|
|
$ p4 delete Main/a
|
|
|
|
|
//depot/Main/a#1 - opened for delete
|
|
|
|
|
$ p4 submit -d second
|
|
|
|
|
Submitting change 2.
|
|
|
|
|
Locking 1 files ...
|
|
|
|
|
delete //depot/Main/a#2
|
|
|
|
|
Change 2 submitted.
|
|
|
|
|
|
|
|
|
|
$ echo a > Main/a
|
|
|
|
|
$ p4 add Main/a
|
|
|
|
|
//depot/Main/a#2 - opened for add
|
|
|
|
|
$ p4 submit -d third
|
|
|
|
|
Submitting change 3.
|
|
|
|
|
Locking 1 files ...
|
|
|
|
|
add //depot/Main/a#3
|
|
|
|
|
Change 3 submitted.
|
|
|
|
|
|
|
|
|
|
Simple import
|
|
|
|
|
|
|
|
|
|
$ cd $hgwd
|
|
|
|
|
$ hg init --config 'format.usefncache=False'
|
2018-06-26 23:28:54 +03:00
|
|
|
$ hg p4seqimport -P $P4ROOT hg-p4-import
|
p4fastimport : introducing fast Perforce to Mercurial convert extension
Summary:
`p4fastimport` is a fast convert extensions for Perforce to Mercurial. It
is designed to generate filelogs in parallel from Perforce. It tries to
minimize the use of Perforce commands and reads from the the Perforce
store on a Perforce server directly.
The core of p4fastimport is the idea to generate a Mercurial filelog
directly from the underlying Perforce data, as a Perforce file in most
cases matches a filelog directly (per-file branches is an exception). To
generate a filelog we are reading each file for an imported revision. A
file in Perforce is locally either stored in RCS, as a compressed GZIP
or as an flat file (binaries). If we do not find a version locally on
disk we fallback to downloading it from Perforce.
We are generating manifests after all filelogs are imported. A manifest
is constructed by adding and removing files from an initial state. We
are generating the correct offset from a manifest into the filelog by
keeping track of how often a file was touched.
We then generate the changelog.
Linkrev generation is a bit tricky. For every file in Perforce know
to which changelist it belongs, as it's stored revisions contains the
changelist. E.g. 1.1422 is the file changed in the changelist 1422 (this
refers to the "original" changelist, before a potential renumbering,
which is why we use the -O switch). We use the CL number obtained
from the revision to reverse lookup the offset in the sorted list of
changelists, which corresponds to it's place in the changelog later,
and therefore it's correct linkrev.
Parallel imports: In order to run parallel imports we MUST keep one lock
at a time, even if we import multiple file logs at the same time. However
filelogs use a singular `fncache`, which will be corrupted if we generate
filelogs in parallel. To avoid this, repositories must be generated with
*fncache* disabled! This restricts `p4fastimport` with workers to run
only on case sensitive file systems.
Test Plan:
The included tests as well as multiple imports from a small testing
Perforce client. Afterwards successfully run `hg verify`
make tests
Reviewers: #idi, quark, durham
Reviewed By: durham
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4776651
Signature: t1:4776651:1492015012:0161c4f45eab4d3b64597d012188c5f2007e8f7d
2017-04-13 21:11:09 +03:00
|
|
|
|
|
|
|
|
Verify
|
|
|
|
|
|
|
|
|
|
$ hg verify
|
|
|
|
|
checking changesets
|
|
|
|
|
checking manifests
|
|
|
|
|
crosschecking files in changesets and manifests
|
|
|
|
|
checking files
|
p4fastimport: handle deleted files + fix wrong parents
Summary:
Running `hg log --traceback --template '{file_copies}' -r XXXX` on a file with long history is slow for 2 reasons
- p4 fast importer preserves full history for deleted and re-added files
- p4 fast importer records the wrong parent of a file
This diff tries to fix these two issues.
In mercurial, if a file is added, deleted, and then added back, it should start a new file history when the file is added again.
For example,
commits commit1 commit2 commit3
actions add a.txt delete a.txt add a.txt
timeline ------------X------------X------------------------X------------
`hg debugindex a.txt` at commit3 shows a.txt as a new file without previous history
rev offset length delta linkrev nodeid p1 p2
0 0 3 -1 0 b789fdd96dc2 000000000000 000000000000
However, this is different in p4. `p4 filelog test.txt` gives you
//depot/Software/Apps/Main/Native/.castle/test.txt
... #3 change 523261 add on 2018/01/23 by zhihuih@devbig415 (text) 'test:add-again-same-file'
... #2 change 523254 delete on 2018/01/23 by zhihuih@devbig415 (text) 'testfile:delete'
... #1 change 523253 add on 2018/01/23 by zhihuih@devbig415 (text) 'testfile:add'
Currently, p4 fast importer preserves history the same way as p4, and this causes slowness (even timeout) in hg when it runs `hg log --traceback --template '{file_copies}' -r XXXX` on a revision that contains files with long history in p4 (mostly contributed by automation). To mitigate this, we want the p4 fast importer to behave the same way as hg, and starts a new history for a file that's added again.
Currently, p4 fast importer takes the tip of a filelog and treats that as the parent of the newly written entry diffusion/FBS/browse/master/fbcode/scm/hg/hgext/p4fastimport/importer.py;19ad9b05f50e3ff0265cdc7b4b45174dcf820343$468-469. This can be wrong when there are revisions from branches.
For example, if I edit file a in master in CL1, 2, 4, and I branch at CL3, and edit the file in branch in CL5, the current importer implementation will take filenode at CL4 as the parent of CL3
(CL1,2,3,4,5 corresponds to rev0,1,3,2,4)
{F120393661}
However, the correct behavior is to take filenode at CL2 as the parent of CL3
(CL1,2,3,4,5 corresponds to rev0,1,3,2,4)
{F120393662}
(This is also the example I use in `test-fb-hgext-p4fastimport-import-branch-filelogorder.t`, so if the description here looks confusing, please refer to the test)
Reviewed By: dsp
Differential Revision: D6962019
fbshipit-source-id: 24de76ae009e0d6f976d247087fe4702c99e0f82
2018-02-22 08:06:40 +03:00
|
|
|
3 files, 3 changesets, 3 total revisions
|
p4fastimport : introducing fast Perforce to Mercurial convert extension
Summary:
`p4fastimport` is a fast convert extensions for Perforce to Mercurial. It
is designed to generate filelogs in parallel from Perforce. It tries to
minimize the use of Perforce commands and reads from the the Perforce
store on a Perforce server directly.
The core of p4fastimport is the idea to generate a Mercurial filelog
directly from the underlying Perforce data, as a Perforce file in most
cases matches a filelog directly (per-file branches is an exception). To
generate a filelog we are reading each file for an imported revision. A
file in Perforce is locally either stored in RCS, as a compressed GZIP
or as an flat file (binaries). If we do not find a version locally on
disk we fallback to downloading it from Perforce.
We are generating manifests after all filelogs are imported. A manifest
is constructed by adding and removing files from an initial state. We
are generating the correct offset from a manifest into the filelog by
keeping track of how often a file was touched.
We then generate the changelog.
Linkrev generation is a bit tricky. For every file in Perforce know
to which changelist it belongs, as it's stored revisions contains the
changelist. E.g. 1.1422 is the file changed in the changelist 1422 (this
refers to the "original" changelist, before a potential renumbering,
which is why we use the -O switch). We use the CL number obtained
from the revision to reverse lookup the offset in the sorted list of
changelists, which corresponds to it's place in the changelog later,
and therefore it's correct linkrev.
Parallel imports: In order to run parallel imports we MUST keep one lock
at a time, even if we import multiple file logs at the same time. However
filelogs use a singular `fncache`, which will be corrupted if we generate
filelogs in parallel. To avoid this, repositories must be generated with
*fncache* disabled! This restricts `p4fastimport` with workers to run
only on case sensitive file systems.
Test Plan:
The included tests as well as multiple imports from a small testing
Perforce client. Afterwards successfully run `hg verify`
make tests
Reviewers: #idi, quark, durham
Reviewed By: durham
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4776651
Signature: t1:4776651:1492015012:0161c4f45eab4d3b64597d012188c5f2007e8f7d
2017-04-13 21:11:09 +03:00
|
|
|
|
|
|
|
|
$ hg update tip
|
|
|
|
|
3 files updated, 0 files merged, 0 files removed, 0 files unresolved
|
|
|
|
|
|
p4fastimport: handle deleted files + fix wrong parents
Summary:
Running `hg log --traceback --template '{file_copies}' -r XXXX` on a file with long history is slow for 2 reasons
- p4 fast importer preserves full history for deleted and re-added files
- p4 fast importer records the wrong parent of a file
This diff tries to fix these two issues.
In mercurial, if a file is added, deleted, and then added back, it should start a new file history when the file is added again.
For example,
commits commit1 commit2 commit3
actions add a.txt delete a.txt add a.txt
timeline ------------X------------X------------------------X------------
`hg debugindex a.txt` at commit3 shows a.txt as a new file without previous history
rev offset length delta linkrev nodeid p1 p2
0 0 3 -1 0 b789fdd96dc2 000000000000 000000000000
However, this is different in p4. `p4 filelog test.txt` gives you
//depot/Software/Apps/Main/Native/.castle/test.txt
... #3 change 523261 add on 2018/01/23 by zhihuih@devbig415 (text) 'test:add-again-same-file'
... #2 change 523254 delete on 2018/01/23 by zhihuih@devbig415 (text) 'testfile:delete'
... #1 change 523253 add on 2018/01/23 by zhihuih@devbig415 (text) 'testfile:add'
Currently, p4 fast importer preserves history the same way as p4, and this causes slowness (even timeout) in hg when it runs `hg log --traceback --template '{file_copies}' -r XXXX` on a revision that contains files with long history in p4 (mostly contributed by automation). To mitigate this, we want the p4 fast importer to behave the same way as hg, and starts a new history for a file that's added again.
Currently, p4 fast importer takes the tip of a filelog and treats that as the parent of the newly written entry diffusion/FBS/browse/master/fbcode/scm/hg/hgext/p4fastimport/importer.py;19ad9b05f50e3ff0265cdc7b4b45174dcf820343$468-469. This can be wrong when there are revisions from branches.
For example, if I edit file a in master in CL1, 2, 4, and I branch at CL3, and edit the file in branch in CL5, the current importer implementation will take filenode at CL4 as the parent of CL3
(CL1,2,3,4,5 corresponds to rev0,1,3,2,4)
{F120393661}
However, the correct behavior is to take filenode at CL2 as the parent of CL3
(CL1,2,3,4,5 corresponds to rev0,1,3,2,4)
{F120393662}
(This is also the example I use in `test-fb-hgext-p4fastimport-import-branch-filelogorder.t`, so if the description here looks confusing, please refer to the test)
Reviewed By: dsp
Differential Revision: D6962019
fbshipit-source-id: 24de76ae009e0d6f976d247087fe4702c99e0f82
2018-02-22 08:06:40 +03:00
|
|
|
Check hg debug data
|
|
|
|
|
$ hg debugdata -m 0
|
|
|
|
|
Main/a\x00b789fdd96dc2f3bd229c1dd8eedf0fc60e2b68e3 (esc)
|
|
|
|
|
Main/b/c\x00149da44f2a4e14f488b7bd4157945a9837408c00 (esc)
|
|
|
|
|
Main/d\x00a9092a3d84a37b9993b5c73576f6de29b7ea50f6 (esc)
|
|
|
|
|
$ hg debugdata -m 1
|
|
|
|
|
Main/b/c\x00149da44f2a4e14f488b7bd4157945a9837408c00 (esc)
|
|
|
|
|
Main/d\x00a9092a3d84a37b9993b5c73576f6de29b7ea50f6 (esc)
|
|
|
|
|
$ hg debugdata -m 2
|
|
|
|
|
Main/a\x00b789fdd96dc2f3bd229c1dd8eedf0fc60e2b68e3 (esc)
|
|
|
|
|
Main/b/c\x00149da44f2a4e14f488b7bd4157945a9837408c00 (esc)
|
|
|
|
|
Main/d\x00a9092a3d84a37b9993b5c73576f6de29b7ea50f6 (esc)
|
|
|
|
|
$ hg debugindex Main/a
|
|
|
|
|
rev offset length delta linkrev nodeid p1 p2
|
|
|
|
|
0 0 3 -1 0 b789fdd96dc2 000000000000 000000000000
|
|
|
|
|
|
p4fastimport : introducing fast Perforce to Mercurial convert extension
Summary:
`p4fastimport` is a fast convert extensions for Perforce to Mercurial. It
is designed to generate filelogs in parallel from Perforce. It tries to
minimize the use of Perforce commands and reads from the the Perforce
store on a Perforce server directly.
The core of p4fastimport is the idea to generate a Mercurial filelog
directly from the underlying Perforce data, as a Perforce file in most
cases matches a filelog directly (per-file branches is an exception). To
generate a filelog we are reading each file for an imported revision. A
file in Perforce is locally either stored in RCS, as a compressed GZIP
or as an flat file (binaries). If we do not find a version locally on
disk we fallback to downloading it from Perforce.
We are generating manifests after all filelogs are imported. A manifest
is constructed by adding and removing files from an initial state. We
are generating the correct offset from a manifest into the filelog by
keeping track of how often a file was touched.
We then generate the changelog.
Linkrev generation is a bit tricky. For every file in Perforce know
to which changelist it belongs, as it's stored revisions contains the
changelist. E.g. 1.1422 is the file changed in the changelist 1422 (this
refers to the "original" changelist, before a potential renumbering,
which is why we use the -O switch). We use the CL number obtained
from the revision to reverse lookup the offset in the sorted list of
changelists, which corresponds to it's place in the changelog later,
and therefore it's correct linkrev.
Parallel imports: In order to run parallel imports we MUST keep one lock
at a time, even if we import multiple file logs at the same time. However
filelogs use a singular `fncache`, which will be corrupted if we generate
filelogs in parallel. To avoid this, repositories must be generated with
*fncache* disabled! This restricts `p4fastimport` with workers to run
only on case sensitive file systems.
Test Plan:
The included tests as well as multiple imports from a small testing
Perforce client. Afterwards successfully run `hg verify`
make tests
Reviewers: #idi, quark, durham
Reviewed By: durham
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4776651
Signature: t1:4776651:1492015012:0161c4f45eab4d3b64597d012188c5f2007e8f7d
2017-04-13 21:11:09 +03:00
|
|
|
End Test
|
|
|
|
|
|
|
|
|
|
stopping the p4 server
|
